Transcriber (STT)
Soniox
Soniox is a real-time STT provider whose stt-rt-v4 model handles 60+ languages — including English and Arabic with native code-switching — and ships semantic end-of-utterance detection built into the model. When you select Soniox, the worker auto-routes turn detection to the STT itself; the platform's standalone turn-detection model is bypassed.
Requires: Soniox API key in Settings → Credentials
Models
| Model | Notes |
|---|---|
stt-rt-v4 | Single unified model. Handles 60+ languages with native code-switching and semantic EOU. |
Configuration
Code
Fields
| Field | Default | Description |
|---|---|---|
model | stt-rt-v4 | The only active Soniox model. |
language_hints | null | ISO codes Soniox should bias toward (e.g. ["en", "ar"]). Omit to let Soniox auto-detect across all supported languages. |
language_hints_strict | false | When true, restrict transcription to only the hinted languages. Pair with language_hints. |
enable_speaker_diarization | false | Identify and label different speakers in the transcript. |
enable_language_identification | true | Tag each token with its detected language — useful for multilingual conversations. |
interim_results | true | Return partial transcripts while the user is speaking. Always-on for Soniox. |
Why pick Soniox over Deepgram
- Multilingual conversations. Soniox's per-token language identification handles English-Arabic code-switching mid-sentence without manual model swapping.
- Semantic EOU. Soniox decides when the speaker has finished a thought based on meaning, not just silence. Often produces lower turn-taking latency on conversational speech without aggressive endpointing tuning.
- Single model, 60+ languages. No per-language model selection.
Why pick Deepgram over Soniox
- Telephony-tuned models (
nova-3,nova-3-medical) and very low best-case latency on English calls. - Per-language formatting controls (smart formatting, numerals, profanity filter).
Conversation behavior interaction
When transcriber.provider = "soniox", the worker sets turn_detection.mode = "stt" automatically — the platform's standalone turn detector is bypassed because Soniox owns endpointing. You can still tune turn_detection.min_delay / max_delay as floor/ceiling guards, but endpointing_mode has no effect.
Related
- Deepgram — telephony-optimised English alternative
- OpenAI STT — Whisper-based fallback
- Phantom STT — Neuratel native STT
Last modified on